/.env.template

###############################################################################
# TIER 1 — QUICK START
# Set this one variable and you're done. Everything else has working defaults.
# Default databases (SQLite, LanceDB, KuzuDB) are file-based, no setup needed.
###############################################################################
LLM_API_KEY="your_api_key"


###############################################################################
# TIER 2 — COMMON OVERRIDES (uncomment to customize)
# Most users only need a few of these.
###############################################################################

# -- LLM Provider & Model ----------------------------------------------------
#LLM_MODEL="openai/gpt-5-mini"
#LLM_PROVIDER="openai"
#LLM_ENDPOINT=""

# -- Embedding Provider -------------------------------------------------------
#EMBEDDING_PROVIDER="openai"
#EMBEDDING_MODEL="openai/text-embedding-3-large"
#EMBEDDING_DIMENSIONS=3072

# -- Tokenizer (chunk sizing) -------------------------------------------------
# The tokenizer used to count tokens for chunking is auto-selected to match the
# embedding model: openai/gemini use TikToken, mistral uses the Mistral
# tokenizer, and fastembed / openai-compatible models use the embedding model's
# own HuggingFace tokenizer. cognee warns (and falls back to TikToken) when it
# cannot match one, since a mismatched tokenizer mis-sizes chunks and skews the
# --dry-run estimate. For providers whose model id is not a HuggingFace repo
# (e.g. Ollama), set HUGGINGFACE_TOKENIZER to a tokenizer matching your model:
#HUGGINGFACE_TOKENIZER="Salesforce/SFR-Embedding-Mistral"

# -- Database Providers (switch from file-based defaults) ---------------------
#DB_PROVIDER="postgres"
#DB_HOST=127.0.0.1
#DB_PORT=5432
#DB_USERNAME=cognee
#DB_PASSWORD=cognee
#DB_NAME=cognee_db

#GRAPH_DATABASE_PROVIDER="neo4j"
#VECTOR_DB_PROVIDER="lancedb"


###############################################################################
# TIER 3 — ADVANCED (grouped by subsystem)
# Most users never need to change anything below this line.
###############################################################################

################################################################################
#  LLM — Advanced Settings
#  Tune these when switching providers, adjusting structured output, or
#  rate-limiting LLM calls.
################################################################################

# Structured output framework: "instructor" (default, via litellm) or "baml"
STRUCTURED_OUTPUT_FRAMEWORK="instructor"

# Instructor's mode determines how structured data is extracted from LLM responses.
# Each LLM has its own default (e.g. gpt-5 models use "json_schema_mode").
#LLM_INSTRUCTOR_MODE=""

# Cognee uses this to determine optimal chunk size (not forwarded in LLM calls).
#LLM_MAX_COMPLETION_TOKENS="16384"

# LLM API version (needed for Azure OpenAI)
#LLM_API_VERSION=""

# Extra kwargs passed to every LLM completion call (JSON string).
# Examples: LLM_ARGS='{"max_tokens": 16384, "temperature": 0.7}'
#LLM_ARGS='{}'

# LLM rate limiting
#LLM_RATE_LIMIT_ENABLED=true
#LLM_RATE_LIMIT_REQUESTS=60
#LLM_RATE_LIMIT_INTERVAL=60

# Per-stage model routing (optional). Unset means the stage uses the base LLM_* config above.
# Route a cheap or local model to extraction (it runs per chunk and dominates token use),
# and keep a stronger model for summarization and query-time reasoning.
#LLM_EXTRACTION_MODEL="ollama_chat/llama3.1"
#LLM_EXTRACTION_PROVIDER="ollama"
#LLM_EXTRACTION_ENDPOINT="http://localhost:11434"
#LLM_EXTRACTION_API_KEY=""
#LLM_SUMMARIZATION_MODEL="openai/gpt-5-mini"
#LLM_SUMMARIZATION_PROVIDER="openai"
#LLM_QUERY_MODEL="openai/gpt-5-mini"
#LLM_QUERY_PROVIDER="openai"

################################################################################
#  Embedding — Advanced Settings
#  Tune these when using non-default embedding providers.
################################################################################

#EMBEDDING_ENDPOINT=""
#EMBEDDING_API_VERSION=""
#EMBEDDING_MAX_COMPLETION_TOKENS=8191
#EMBEDDING_BATCH_SIZE=36
# If not provided, LLM_API_KEY is used for embeddings too.
#EMBEDDING_API_KEY="your_api_key"

################################################################################
#  BAML Structured Output
#  Only needed when STRUCTURED_OUTPUT_FRAMEWORK="baml".
################################################################################

#BAML_LLM_PROVIDER=openai
#BAML_LLM_MODEL="gpt-5-mini"
#BAML_LLM_ENDPOINT=""
#BAML_LLM_API_KEY="your_api_key"
#BAML_LLM_API_VERSION=""

################################################################################
#  Root Directories
#  Override where Cognee stores files and databases (default: .venv).
################################################################################

#DATA_ROOT_DIRECTORY='/Users/<user>/Desktop/cognee/.cognee_data/'
#SYSTEM_ROOT_DIRECTORY='/Users/<user>/Desktop/cognee/.cognee_system/'

################################################################################
#  Storage Backend
#  Switch from local filesystem to S3.
################################################################################

#STORAGE_BACKEND="local"
#STORAGE_BACKEND="s3"
#STORAGE_BUCKET_NAME="your-bucket-name"
#AWS_REGION="us-east-1"
#AWS_ACCESS_KEY_ID="your-access-key"
#AWS_SECRET_ACCESS_KEY="your-secret-key"
#DATA_ROOT_DIRECTORY="s3://your-bucket/cognee/data"
#SYSTEM_ROOT_DIRECTORY="s3://your-bucket/cognee/system"
#CACHE_ROOT_DIRECTORY="s3://your-bucket/cognee/cache"

################################################################################
#  Relational Database — Advanced
#  Connection tuning, pool sizes, SSL.
################################################################################

DB_PROVIDER="sqlite"
DB_NAME=cognee_db

# Custom connection arguments (JSON). Useful for SSL, timeouts.
#DATABASE_CONNECT_ARGS='{"sslmode": "require", "connect_timeout": 10}'

# Connection pool tuning (JSON).
#POOL_ARGS='{"pool_size": 5, "max_overflow": 10, "pool_recycle": -1, "pool_timeout": 30}'

################################################################################
#  Graph Database — Advanced
#  Provider-specific connection details.
################################################################################

GRAPH_DATABASE_PROVIDER="kuzu"
# Handler for multi-user access control (per-dataset DB creation).
GRAPH_DATASET_DATABASE_HANDLER="kuzu"

# Remote Kuzu
#GRAPH_DATABASE_PROVIDER="kuzu-remote"
#GRAPH_DATABASE_URL="http://localhost:8000"
#GRAPH_DATABASE_USERNAME=XXX
#GRAPH_DATABASE_PASSWORD=YYY

# Neo4j
#GRAPH_DATABASE_PROVIDER="neo4j"
#GRAPH_DATABASE_URL=bolt://localhost:7687
#GRAPH_DATABASE_NAME="neo4j"
#GRAPH_DATABASE_USERNAME=neo4j
#GRAPH_DATABASE_PASSWORD=pleaseletmein

################################################################################
#  Vector Database — Advanced
#  Provider-specific connection details.
################################################################################

# Supported (built-in): pgvector | lancedb
# Community adapters (separate packages): qdrant | weaviate | milvus | chromadb
VECTOR_DB_PROVIDER="lancedb"
#VECTOR_DB_URL=
#VECTOR_DB_KEY=
# Handler for multi-user access control (per-dataset DB creation).
VECTOR_DATASET_DATABASE_HANDLER="lancedb"

# Connection pool tuning for PGVector per-dataset engines (JSON).
# When ENABLE_BACKEND_ACCESS_CONTROL=true each dataset gets its own engine; this controls
# its pool size independently from POOL_ARGS (default: pool_size=2, max_overflow=2).
#VECTOR_POOL_ARGS='{"pool_size": 2, "max_overflow": 5, "pool_recycle": 1800}'

################################################################################
#  Ontology Resolver
#  Use when grounding extraction against an OWL ontology.
################################################################################

#ONTOLOGY_RESOLVER=rdflib
#MATCHING_STRATEGY=fuzzy
#ONTOLOGY_FILE_PATH=YOUR_FULL_FILE_PATH

################################################################################
#  Database Adapter Caching
#  Max graph / vector / relational engine instances held in the LRU cache
#  (one per unique connection key, e.g. per dataset in multi-tenant mode).
#  In subprocess mode, this also caps how many child processes (Kuzu/LanceDB
#  workers) can be alive at once — eviction shuts down the subprocess.
#  Also the default for DATASET_QUEUE_MAX_CONCURRENT when that is unset.
#  Engines of datasets currently admitted by the dataset queue are pinned and
#  never evicted by capacity pressure; when every entry is pinned the cache
#  briefly exceeds this size (bounded by DATASET_QUEUE_MAX_CONCURRENT).
#  Lower values save memory; raise when running many datasets concurrently.
################################################################################

#DATABASE_MAX_LRU_CACHE_SIZE=6

################################################################################
#  Dataset Queue
#  Semaphore-backed queue that limits how many datasets can be processed at
#  once (cognify, search, etc.). Prevents resource exhaustion when many
#  datasets run in parallel. When the limit is reached, new datasets wait
#  until a slot is freed.
################################################################################

#DATASET_QUEUE_ENABLED=true
# Max concurrent dataset slots. Defaults to DATABASE_MAX_LRU_CACHE_SIZE.
#DATASET_QUEUE_MAX_CONCURRENT=6

################################################################################
#  Translation
#  Use when ingesting non-English content.
################################################################################

TRANSLATION_PROVIDER="llm"
TARGET_LANGUAGE="en"
CONFIDENCE_THRESHOLD=0.8
#GOOGLE_TRANSLATE_API_KEY="your-google-api-key"
#GOOGLE_PROJECT_ID="your-google-project-id"
#AZURE_TRANSLATOR_KEY="your-azure-translator-key"
#AZURE_TRANSLATOR_REGION="westeurope"
#AZURE_TRANSLATOR_ENDPOINT="https://api.cognitive.microsofttranslator.com"
#TRANSLATION_BATCH_SIZE=10
#TRANSLATION_MAX_RETRIES=3
#TRANSLATION_TIMEOUT_SECONDS=30

################################################################################
#  Data Migrations (graph/vector revision chain)
################################################################################

# Cognee runs its data migrations automatically on startup (FastAPI lifespan,
# first remember()/cognify() call in an SDK process). Set to false to disable
# ALL automatic runs and migrate explicitly via `cognee-cli upgrade` instead
# (e.g. operator-driven deployments, or tests on deliberately old-format data).
#ENABLE_AUTO_MIGRATIONS=true

################################################################################
#  Migration (Relational -> Graph)
################################################################################

MIGRATION_DB_PATH="/path/to/migration/directory"
MIGRATION_DB_NAME="migration_database.sqlite"
MIGRATION_DB_PROVIDER="sqlite"
#MIGRATION_DB_USERNAME=cognee
#MIGRATION_DB_PASSWORD=cognee
#MIGRATION_DB_HOST="127.0.0.1"
#MIGRATION_DB_PORT=5432

################################################################################
#  Security
################################################################################

# -- JWT Authentication -------------------------------------------------------
# Secret used to sign and verify JWT tokens. Must be the same across all instances
# (e.g. all Kubernetes pods) for tokens issued by one instance to be accepted by another.
# Change this to a long random string in production. Never commit the real value to git.
FASTAPI_USERS_JWT_SECRET="super_secret"

# How long a JWT token remains valid, in seconds. After expiry the user must log in again.
# The same lifetime applies to both cookie and bearer token auth.
# Default: 3600 (1 hour)
JWT_LIFETIME_SECONDS=3600

# -- API Key Authentication ---------------------------------------------------
# When HASH_API_KEY=true, API keys are hashed with SHA-256 before being stored in the database.
# This means the raw key is shown to the user only once at creation time and cannot be recovered.
#
# ⚠️  Migration note: if you enable this on a running system that already has API keys stored
#     in plaintext, those existing keys will stop working immediately because the lookup will
#     hash the incoming value and find no match. You must either:
#       1. Delete and re-issue all existing API keys, or
#       2. Write a one-off migration to SHA-256 hash the existing api_key column values.
#
# Default: false (keys are stored in plaintext)
HASH_API_KEY="False"

# When set to false don't allow adding of local system files to Cognee. Should be set to False when Cognee is used as a backend.
ACCEPT_LOCAL_FILE_PATH=True
ALLOW_HTTP_REQUESTS=True
ALLOW_CYPHER_QUERY=True
RAISE_INCREMENTAL_LOADING_ERRORS=True

# Authentication & access control.
#
# ENABLE_BACKEND_ACCESS_CONTROL is the canonical posture switch:
#   true  (default) - multi-tenant mode: per-user/dataset isolated DBs AND
#                     API endpoints require an authenticated user.
#   false           - single-user mode: shared DB AND auth requirement off.
#
# REQUIRE_AUTHENTICATION is an explicit override on the auth requirement only:
#   unset (default) - follow ENABLE_BACKEND_ACCESS_CONTROL.
#   true            - force auth on (sane for single-user behind a token).
#   false           - force auth off — IGNORED if ENABLE_BACKEND_ACCESS_CONTROL
#                     is true (multi-tenant always requires auth; a warning is
#                     logged at startup).
#
# Startup logs an "auth posture: ..." line with the resolved decision so you
# can verify what's actually in effect.
REQUIRE_AUTHENTICATION=False

# Set this variable to True to enforce usage of backend access control for Cognee
# Note: This is only currently supported by the following databases:
#       Relational: SQLite, Postgres
#       Vector: LanceDB, pgvector
#       Graph: KuzuDB, neo4j_aura_dev
#
# It enforces creation of databases per Cognee user + dataset. Does not work with some graph and database providers.
# Disable mode when using not supported graph/vector databases.
ENABLE_BACKEND_ACCESS_CONTROL=True

################################################################################
#  Cloud Sync
################################################################################

COGNEE_CLOUD_API_URL="http://localhost:8001"
COGNEE_CLOUD_AUTH_TOKEN="your-api-key"

################################################################################
#  UI
################################################################################

UI_APP_URL=http://localhost:3000

################################################################################
#  DLT Ingestion
################################################################################

#DLT_MAX_ROWS_PER_TABLE=50

################################################################################
#  Dev / Debug
################################################################################

ENV="local"
#ENABLE_LAST_ACCESSED="false"
TOKENIZERS_PARALLELISM="false"

# -- Search History ------------------------------------------------------------
# Set to false to disable search query/result logging (recommended for daemons)
#COGNEE_LOG_SEARCH_HISTORY="true"

# LITELLM Logging Level. Set to quiet down logging
LITELLM_LOG="ERROR"
#TELEMETRY_DISABLED=1
#DEFAULT_USER_EMAIL=""
#DEFAULT_USER_PASSWORD=""

# -- Cognee Logging -----------------------------------------------------------
# Console log level: DEBUG, INFO, WARNING, ERROR, CRITICAL (default: INFO)
#LOG_LEVEL="INFO"
# Set to false to disable file logging entirely (console-only)
#COGNEE_LOG_FILE="true"
# Override the log directory (default: ~/.cognee/logs)
#COGNEE_LOGS_DIR="/var/log/cognee"
# Max size per log file before rotation, in bytes (default: 50 MB)
#COGNEE_LOG_MAX_BYTES=52428800
# Number of rotated log files to keep (default: 5 → 300 MB total cap)
#COGNEE_LOG_BACKUP_COUNT=5

################################################################################
#  AWS
################################################################################

#AWS_REGION=""
#AWS_ENDPOINT_URL=""
#AWS_ACCESS_KEY_ID=""
#AWS_SECRET_ACCESS_KEY=""
#AWS_SESSION_TOKEN=""

################################################################################
#  Web Scraper
################################################################################

WEB_SCRAPER_TIMEOUT=15.0
WEB_SCRAPER_MAX_DELAY=10.0

################################################################################
#  OpenTelemetry / Tracing
################################################################################

# -- To export traces to an OTLP-compatible backend (Dash0, Grafana, Jaeger, etc.),
#    set the endpoint and optional auth headers: ---------------------
# COGNEE_TRACING_ENABLED=true
# OTEL_EXPORTER_OTLP_ENDPOINT="https://ingress.eu-west.dash0.com:4317"
# OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer <your-token>"

# Override the service name reported in traces (default: "cognee")
# OTEL_SERVICE_NAME="cognee"

# Add extra resource attributes (useful for Kubernetes, multi-instance deployments)
# OTEL_RESOURCE_ATTRIBUTES="service.namespace=my-team,service.version=1.0"

# Session cache settings
# To switch to Redis caching check our documentation page sessions-and-caching
# CACHING=true
# Backends: sqlite (default), postgres, redis, fs, tapes
# CACHE_BACKEND=sqlite
# CACHE_BACKEND=postgres
# Optional explicit SQLAlchemy async URL for the sqlite/postgres backends.
# sqlite default: cache.db next to the relational SQLite database.
# postgres default: falls back to DB_* settings when DB_PROVIDER=postgres.
# CACHE_DB_URL=sqlite+aiosqlite:///path/to/databases/cache.db
# CACHE_DB_URL=postgresql+asyncpg://cognee:cognee@localhost:5432/cognee_db
# Minimum seconds between global TTL purge sweeps (sqlite/postgres backends)
# CACHE_PURGE_INTERVAL_SECONDS=900


################################################################################
#  ADDITIONAL MANAGED SETTINGS (previously undocumented)
#  These are all read by Cognee's config classes (pydantic BaseSettings) but
#  were missing from this template. Defaults shown; uncomment to override.
################################################################################

# -- LLM tuning ---------------------------------------------------------------
#LLM_TEMPERATURE=0.0
#LLM_STREAMING=false
# Optional fallback model used when the primary completion fails.
#FALLBACK_MODEL=""
#FALLBACK_API_KEY=""
#FALLBACK_ENDPOINT=""
# Audio transcription model.
#TRANSCRIPTION_MODEL="whisper-1"

# -- Embedding rate limiting (mirrors the LLM_RATE_LIMIT_* knobs) -------------
#EMBEDDING_RATE_LIMIT_ENABLED=false
#EMBEDDING_RATE_LIMIT_REQUESTS=60
#EMBEDDING_RATE_LIMIT_INTERVAL=60
#EMBEDDING_RATE_LIMIT_TOKENS=0
# Token-based LLM limit (0 = disabled; requests/interval already documented above).
#LLM_RATE_LIMIT_TOKENS=0

# -- Chunking -----------------------------------------------------------------
#CHUNK_SIZE=1500
#CHUNK_OVERLAP=10
#CHUNK_STRATEGY="paragraph"

# -- Triplet embedding (extra triplet-level vectors during cognify) -----------
#TRIPLET_EMBEDDING=false

# -- Session cache (Redis backend + session/usage tuning) ---------------------
# Used when CACHE_BACKEND=redis; also the host/port for a remote cache.
#CACHE_HOST="localhost"
#CACHE_PORT=6379
#CACHE_USERNAME=""
#CACHE_PASSWORD=""
# Session lifetime in the cache (default 7 days) and per-turn context cap.
#SESSION_TTL_SECONDS=604800
#MAX_SESSION_CONTEXT_CHARS=
# Self-improvement: absorb per-turn feedback/guidance automatically (default on).
#AUTO_FEEDBACK=true
# Per-process LLM usage logging into the cache.
#USAGE_LOGGING=false
#USAGE_LOGGING_TTL=604800
# Cross-process locks for file-based embedded graph backends.
#SHARED_KUZU_LOCK=false
#SHARED_LADYBUG_LOCK=false

# -- Graph database — advanced connection / Kuzu tuning -----------------------
#GRAPH_DATABASE_HOST=""
#GRAPH_DATABASE_PORT=
#GRAPH_DATABASE_KEY=""
#GRAPH_DATABASE_ALLOW_ANONYMOUS=false
# Run the embedded graph engine (Kuzu/Ladybug) in a worker subprocess.
#GRAPH_DATABASE_SUBPROCESS_ENABLED=true
# Kuzu performance tuning (0/auto by default).
#KUZU_NUM_THREADS=0
#KUZU_BUFFER_POOL_SIZE=
#KUZU_MAX_DB_SIZE=

# -- Vector database — advanced connection ------------------------------------
#VECTOR_DB_HOST=""
#VECTOR_DB_PORT=1234
#VECTOR_DB_NAME=""
#VECTOR_DB_USERNAME=""
#VECTOR_DB_PASSWORD=""
#VECTOR_DB_SUBPROCESS_ENABLED=true

# -- Database subprocess workers — advanced tuning ----------------------------
# The embedded DB engines (Kuzu/Ladybug graph, LanceDB vector) run their native
# client in a dedicated worker process. These knobs tune that harness.
# Per-RPC deadline guarding against a hung native call (seconds; <=0 disables).
#SUBPROCESS_CALL_TIMEOUT=300
# How many times a failed subprocess RPC is retried (respawning the worker).
#SUBPROCESS_MAX_RETRIES=2
# Backstop for the brief window where one graph worker is still releasing a
# file lock while another opens the same DB path: the worker retries the open
# this many times, with exponential backoff starting at this many seconds
# (per-attempt backoff is capped internally).
#SUBPROCESS_OPEN_LOCK_RETRIES=10
#SUBPROCESS_OPEN_LOCK_BACKOFF=0.1

# -- AWS / Bedrock extras (in addition to the AWS section above) --------------
#AWS_PROFILE_NAME=""
#AWS_BEDROCK_RUNTIME_ENDPOINT=""

# -- Local llama.cpp provider -------------------------------------------------
#LLAMA_CPP_MODEL_PATH=""
#LLAMA_CPP_N_CTX=2048
#LLAMA_CPP_N_GPU_LAYERS=0
#LLAMA_CPP_CHAT_FORMAT="chatml"

# -- Security: additional auth-token secrets ----------------------------------
# Like FASTAPI_USERS_JWT_SECRET above, these default to the INSECURE value
# "super_secret". Override BOTH with long random strings in production.
#FASTAPI_USERS_VERIFICATION_TOKEN_SECRET="change_me_in_production"
#FASTAPI_USERS_RESET_PASSWORD_TOKEN_SECRET="change_me_in_production"


################################################################################
#  Docker / MCP Runtime
#  Configure the cognee API image (cognee/cognee) and the MCP image
#  (cognee/cognee-mcp) when running `docker run` / `docker compose`.
#  Unless noted "read by the app", these are consumed by the container
#  entrypoints/compose and have defaults baked into the images — set them only
#  to override. (docker-compose.yml already sets sensible values for most.)
################################################################################

# -- API server (cognee/cognee image) ----------------------------------------
# CORS allow-list for the FastAPI server: comma-separated origins. Read by the
# app (cognee/api/client.py). Default '*' (all origins) — set explicit domains
# in production.
#CORS_ALLOWED_ORIGINS="https://yourdomain.com,https://another.com"
# Server bind/port inside the container (entrypoint defaults shown).
#HTTP_PORT=8000
#BIND_ADDRESS=0.0.0.0

# -- MCP server (cognee/cognee-mcp image) -------------------------------------
# Transport the MCP container serves. The Docker image reads TRANSPORT_MODE;
# the direct `cognee-mcp` CLI uses --transport instead.
#TRANSPORT_MODE=stdio          # stdio | sse | http
# Comma-separated optional extras to pip-install at container startup.
#EXTRAS=aws,postgres
# MCP "API mode": point the MCP server at an already-running cognee API server.
#API_URL=http://localhost:8000
#API_TOKEN=""
# MCP "Cloud mode": point the MCP server at a managed Cognee Cloud instance.
# These are the canonical cloud-connection variables, shared across serve(),
# push(), the MCP server, and sync. COGNEE_CLOUD_API_URL / COGNEE_CLOUD_AUTH_TOKEN
# (above) remain as deprecated fallbacks.
#COGNEE_SERVICE_URL=""
#COGNEE_API_KEY=""

# -- Debug (both images) ------------------------------------------------------
# DEBUG=true together with ENV in {dev,local} starts the container under
# debugpy, listening on DEBUG_PORT. ENV is the canonical environment variable
# (set it in the Dev/Debug section above); ENVIRONMENT is a deprecated alias
# still accepted by the container entrypoints.
#DEBUG=false
#DEBUG_PORT=5678

# -- Frontend (cognee-frontend image / compose `ui` profile) ------------------
#NEXT_PUBLIC_BACKEND_API_URL=http://localhost:8000


###############################################################################
# TIER 4 — EXAMPLE PROVIDER OVERRIDES (commented out)
# Uncomment + fill values to switch providers.
###############################################################################

########## Azure OpenAI (API key auth) ########################################
#LLM_PROVIDER="azure"
#LLM_MODEL="azure/gpt-5-mini"
#LLM_ENDPOINT="https://YOUR-RESOURCE.openai.azure.com"
#LLM_API_KEY="your-azure-api-key"
#LLM_API_VERSION="2024-12-01-preview"
#LLM_MAX_COMPLETION_TOKENS="16384"

########## Azure OpenAI (managed identity / DefaultAzureCredential) ###########
# Uses DefaultAzureCredential - no API key needed (for Azure VMs, App Service, etc.)
# Requires: pip install azure-identity
#LLM_PROVIDER="azure"
#LLM_MODEL="azure/gpt-5-mini"
#LLM_ENDPOINT="https://YOUR-RESOURCE.openai.azure.com"
#LLM_API_VERSION="2024-12-01-preview"
#LLM_AZURE_USE_MANAGED_IDENTITY=true

#EMBEDDING_MODEL="azure/text-embedding-3-large"
#EMBEDDING_ENDPOINT="https://YOUR-RESOURCE.openai.azure.com/openai/deployments/text-embedding-3-large"
#EMBEDDING_API_KEY="your-azure-api-key"
#EMBEDDING_API_VERSION="2024-12-01-preview"
#EMBEDDING_DIMENSIONS=3072
#EMBEDDING_MAX_COMPLETION_TOKENS=8191

########## Local LLM via Ollama ###############################################
#LLM_API_KEY ="ollama"
#LLM_MODEL="llama3.1:8b"
#LLM_PROVIDER="ollama"
#LLM_ENDPOINT="http://localhost:11434/v1"
#EMBEDDING_PROVIDER="ollama"
#EMBEDDING_MODEL="nomic-embed-text:latest"
#EMBEDDING_ENDPOINT="http://localhost:11434/api/embed"
#EMBEDDING_DIMENSIONS=768
#HUGGINGFACE_TOKENIZER="nomic-ai/nomic-embed-text-v1.5"

########## OpenRouter (also free) #############################################
#LLM_API_KEY="<<go-get-one-yourself"
#LLM_PROVIDER="custom"
#LLM_MODEL="openrouter/google/gemini-2.0-flash-lite-preview-02-05:free"
#LLM_ENDPOINT="https://openrouter.ai/api/v1"

########## DeepInfra ##########################################################
#LLM_API_KEY="<<>>"
#LLM_PROVIDER="custom"
#LLM_MODEL="deepinfra/meta-llama/Meta-Llama-3-8B-Instruct"
#LLM_ENDPOINT="https://api.deepinfra.com/v1/openai"
#EMBEDDING_PROVIDER="openai"
#EMBEDDING_API_KEY="<<>>"
#EMBEDDING_MODEL="deepinfra/BAAI/bge-base-en-v1.5"
#EMBEDDING_ENDPOINT=""
#EMBEDDING_API_VERSION=""
#EMBEDDING_DIMENSIONS=3072
#EMBEDDING_MAX_COMPLETION_TOKENS=8191